A new AI system revises rejected research papers by looping them through an AI reviewer - and it's much better at fixing bugs than fixing bad ideas.
Researchers built AppliedScientist, which pairs an autonomous AI scientist that edits paper drafts with an AI reviewer that grades each revision independently, with no memory of earlier scores, so feedback doesn't compound its own bias. They tested three ways to start the loop: seeding it with a paper's original human reviews, seeding it with AI-generated reviews, or letting the system critique its own drafts with a fixed prompt every round. Reviewer-guided revision beat blind self-revision, and an independent evaluator called Stanford Reviewer confirmed the later drafts scored higher too. Out of 150 flagged weaknesses in methodology or execution, the system fixed 128 of them - 85.3% - but of 18 weaknesses about the paper's core idea, it fixed just two.
That split is the real finding. AI can tighten sloppy experiments and patch implementation gaps, the kind of feedback a sharp lab mate gives for free. It cannot reliably tell a researcher their idea isn't interesting enough, which is usually the harder and more valuable half of peer review.
Every automated-reviewing tool so far has been decent at catching weak ablations and missing baselines; none of them can yet tell you your premise is boring, and that gap looks like the part that actually matters.